Selection of future-query horizon sets

Determine which future-query sets should be used to train the latent world model and planner.

Background

The paper evaluates several future-query configurations, including single horizons at 1 or 2 seconds, the pair {1,2}, and the four-horizon set {1,2,3,4}. Four horizons achieve the highest reported PDMS both without and with Trajectory-Adaptive Bias, but the performance differences are small, and horizon-shuffling experiments suggest that the planner uses the predictions primarily as aggregated multi-scale temporal context rather than as a strictly ordered rollout.

Although the experiments compare several configurations, they do not establish which horizon set is generally preferable or whether the observed advantage of four horizons is robust. The unresolved issue concerns the appropriate future-query design for training the latent world model and planner.